Jason Liu details OpenAI Codex workflow for persistent knowledge work
New integration strategies allow AI agents to maintain context across sessions, automate routine monitoring, and produce durable digital artifacts without constant human intervention.
Jason Liu has published a detailed account of his workflow integration with OpenAI Codex, outlining a strategic shift from utilising coding agents for software development to employing them for broader knowledge work. In a post titled "Codex-maxxing" published on 10 May 2026, Liu describes a system where AI agents manage long-running workstreams through persistent "pinned threads," voice input, and shared memory architectures. The approach moves beyond simple prompt-and-response interactions, creating what Liu terms an "operating loop" where work continues autonomously between user sessions.
Central to this workflow is the use of compaction in long-running threads to maintain context over months. Liu accepts the higher computational costs associated with these megathreads, arguing that continuity is essential for preserving history, preferences, and decisions. To capture unedited thinking, he utilises voice input via Wispr Flow for system-wide dictation and Granola for recording conversations, feeding these transcripts directly into Codex. This allows the agent to process "messy" context that might be too cumbersome to type, enabling more natural steering of tasks while the agent is still executing tool calls.
Liu distinguishes between OpenAI’s native personalisation memories and a custom, durable memory system built on an Obsidian vault stored as a GitHub repository. This file-based memory allows agents to serialize insights into inspectable artifacts, ensuring that knowledge survives if a thread dies or becomes too expensive to maintain. By reviewing diffs in the vault, Liu can verify what the agent deemed important, transforming the system from a chat interface into a shared notebook where different workers read from the same source of truth.
Automation is driven by "Heartbeats," which enable recurring monitoring of tasks at set intervals. Liu cites examples such as a "Chief of Staff" thread that drafts Slack replies every 30 minutes and an animation project where a heartbeat monitored feedback every 15 minutes, triggered a re-render, and posted the file via desktop automation tools. The workflow integrates connectors for Slack, Gmail, and Amazon, allowing the agent to navigate authenticated browser sessions and desktop applications to execute complex tasks, such as securing a refund in approximately 25 minutes.
The interface for this work is the Codex side panel, which Liu utilises to inspect and annotate artifacts such as HTML, spreadsheets, and presentations without breaking the operational loop. He expresses a preference for single-file HTML outputs over Markdown, viewing them as durable, interactive applications that can be updated via Heartbeats. By combining persistent threads, voice input, and automated monitoring, Liu argues that the value of AI agents lies not just in code generation, but in their ability to keep work moving forward autonomously.
